demand Fluctuations in the Ready-Mix Concrete...
Transcript of demand Fluctuations in the Ready-Mix Concrete...
Demand Fluctuations in the Ready-Mix ConcreteIndustry
Allan Collard-Wexler∗
NYU Stern and NBER
November 29, 2012
Abstract
I investigate the role of demand shocks in the ready-mix concrete industry. UsingCensus data on more than 15,000 plants, I estimate a model of investment and entryin oligopolistic markets. These estimates are used to simulate the effect of eliminatingshort-term local demand changes. A policy of smoothing the volatility of demand has amarket expansion effect: The model predicts a 39% increase in the number of plants inthe industry. Since bigger markets have both more plants and larger plants, a demand-smoothing fiscal policy would increase the share of large plants by 20%. Finally, thepolicy of smoothing demand reduces entry and exit by 25%, but has no effect on therate at which firms change their size.
∗The work in this paper is drawn from chapter 2 of my Ph.D. dissertation at Northwestern Universityunder the supervision of Mike Whinston, Rob Porter, Shane Greenstein, and Aviv Nevo. I would like tothank the anonymous referees for comments that greatly improved the paper, as well as John Asker, LanierBenkard, Ambarish Chandra, Alessandro Gavazza, Mike Mazzeo, Ariel Pakes, Lynn Riggs, and Stan Zinfor helpful conversations. The Fonds Quebecois de la Recherche sur la Societe et la Culture (FQRSC) andthe Center for the Study of Industrial Organization at Northwestern University (CSIO) provided financialsupport. I would like to thank seminar participants at many institutions for comments. The research inthis paper was conducted while I was a Special Sworn Status researcher of the U.S. Census Bureau at theChicago Census Research Data Center. Research results and conclusions expressed are those of the authorand do not necessarily reflect the views of the Census Bureau. All results have been reviewed to ensure thatno confidential information is disclosed. Support for this research at the Chicago and New York RDC fromthe NSF (awards no. SES-0004335 and ITR-0427889) is also gratefully acknowledged.
1
1 Introduction
Many industries face considerable uncertainty about future demand for their products. How
do these shocks affect the organization of production?
I study the effect of demand shocks in the ready-mix concrete industry. This industry is
composed of local oligopolies as wet concrete cannot travel much more than an hour before
hardening. The ready-mix concrete industry experiences large changes in demand from the
construction sector from year-to-year, as the size of the local construction industry fluctuates
by an average of 30% per year. Moreover, about half of all concrete is purchased by state and
local governments, and these outlays are particularly volatile, due to year-to-year variation
in tax revenues.
To investigate the role of demand volatility, I estimate a model of entry and discrete in-
vestment in concentrated markets using an Indirect Inference Conditional Choice Probability
Algorithm, which allows for considerable plant heterogeneity.1 This model is estimated with
Census data on the histories of more than 15,000 ready-mix concrete plants in the United
States from 1976 to 1999.
Plant size is directly related to market size in the ready-mix concrete sector: Bigger
markets have both more plants and larger plants. Thus the model’s estimates show that
construction employment has strong positive effects on profits but disproportionately affects
large plants. Competition – in particular, the presence of a first competitor – substantially
reduces profits. Firms pay large sunk costs both for entering the market and for increasing
or shrinking the size of a plant.
I look at government intervention in the ready-mix concrete market that smooths out
short-term fluctuations in demand at the county level. Specifically, the counterfactual mimics
1Previous versions of the paper used an algorithm analogous to Aguiregabiria and Mira (2007) where thechoice probabilities were updated to match those given by a computed equilibrium of the game, given theestimated parameter vector. This technique leads to similar estimates and counterfactual results as thosepresented in the paper and are available by request.
2
the effect of government sequencing its contracts so as to spread demand evenly over each
five year period. However, secular changes in demand, those that move average demand from
one five year period to the next, are preserved. Thus this policy eliminates short-run – i.e.
five year – changes in demand, but preserves longer run movement in demand.
I find that this demand smoothing policy reduces entry and exit rates, by 25%, but has
no effect on the rate at which plants change their size. The modest effect of the demand
smoothing policy on the dynamics of the industry is due to high estimates of sunk costs of
both entry and adjustment. These make it costly for firms to react to short-lived changes in
demand. In addition, when demand becomes less volatile, firms get a more precise forecast
of future demand. Thus the direct effect of a smoother demand process is offset by firms
becoming more responsive to the remaining changes in demand. This lessens any effect of
demand smoothing on turnover.
However, smoothing demand also has a large “market expansion” effect – it raises the
number of plants in the industry by 39%. The inter-temporal volatility of demand can have
large effects on the profitability of a market. To illustrate, consider that in the market
for electricity, demand volatility is thought to raise the profits of generators. In periods of
peak demand, capacity constraints bind and spot prices can increase quite dramatically.
Alternatively, in the ready-mix concrete market, periods of peak demand might raise costs
due to the congestion associated with multiple concrete deliveries. In the data, a one percent
increase in market size (as measured by construction employment) is associated with a 0.69%
increase in the number of ready-mix concrete plants. This indicates a concave response to
higher demand, and these non-linearities of period profits, with respect to demand, indicate
that demand volatility affects market size. Thus some of the more interesting effects of the
demand smoothing policy are expressed in the industry’s cross-section rather than in it’s
year-to-year changes.
The counterfactual of smoothing demand raises investment by 44%, from $439 million
3
to $634 million per year, but decreases producer surplus for incumbents by 20%. Moreover,
since the market expansion effect is similar to an increase in market size, the size distribution
of the industry shifts toward large plants, and the share of large plants (with more than 17
employees) climbs by 20%. Turning towards the effect on consumers, the 39% increase in the
number of plants due to the demand smoothing policy would reduce the share of monopoly
markets from 43 to 25 percent. The resulting increase in competition would make prices fall,
and consumers would pay $43 million less per year for ready-mix concrete.
The effect of counter-cyclical fiscal policy – in particular as regards to the response,
timing and composition of investments – has been extensively discussed in the public finance
literature (see Auerbach, Gale, and Harris (2010) and the references therein). This paper
casts light on the the effect of active fiscal policy not only on the dynamics of the industry
in terms of entry and exit, but also on market structure and industry composition.
This paper proceeds as follows. In section 2, I discuss the ready-mix concrete industry.
Section 3 describes the data. In section 4, I present a dynamic model of competition. I
describe estimation in section 5 and results in section 6. Finally, in section 7, I analyze the
effect of policies that would eliminate some of the volatility of demand.
2 The Ready-Mix Concrete Industry
2.1 The Industry
2.1.1 Concrete
I focus on ready-mix concrete: concrete mixed with water at a plant and transported directly
to a construction site.2 While it is possible to produce several hundred types of concrete,
2Concrete is a mixture of three basic ingredients: sand, gravel (crushed stone), and cement, as well aschemical compounds known as admixtures. Combining this mixture with water turns cement into a hardpaste that binds the sand and gravel together.
4
these mixtures basically use the same ingredients and machinery. Thus one can think of
ready-mix concrete as a homogeneous product.
Ready-mix is a perishable product that must be delivered within an hour and a half
before it becomes too stiff to be workable.3,4 Concrete is also very cheap for its weight, and
so ready-mix trucks typically drive 20 minutes to deliver their loads.5
2.1.2 Local Oligopoly
Due to these high transportation costs, concrete markets are geographically segmented: Fig-
ure 1 shows the dispersion of ready-mix producers in the Midwest, with a handful of in-
cumbents in each area. For my empirical work, I treat each county as a separate market
that evolves independently from the rest of the industry. Furthermore, Table 1 shows that
the vast majority of counties in the United States have fewer than six ready-mix plants,
reflecting a locally oligopolistic market structure. However, because even the most isolated
rural areas have some demand for ready-mix concrete, most counties are served by at least
one producer.
A market with more than three firms appears to yield fairly competitive outcomes. To
illustrate, Figure 2 shows the median price of ready-mix concrete in markets with one to
seven firms.6 The first three competitors have a noticeable effect on prices, but additional
3One producer describes the economics of transportation costs in the ready-mix industry as follows:
“A truckload of concrete contains about 7 cubic yards of concrete. A cubic yard of concreteweighs about 4,000 pounds and will cost you around $60 delivered to your door. That’s 1.5cents a pound. If you go to your local hardware store, you get a bag of manure weighing 10pounds for $5. That means that concrete is cheaper than shit.”
4“ASTM C 94 also requires that concrete be delivered and discharged within 1 1/2 hours or before thedrum has revolved 300 times after the introduction of water to the cement and aggregates” p.96 in Kosmatka,Kerkhoff, and Panarese (2002).
5The average price of concrete is around 1.5 cents per pound. The driving time of twenty minutes is basedon a dozen interviews conducted with Illinois ready-mix concrete producers. Thanks to Dick Plimpton atthe Illinois Ready-Mix Concrete Association for providing IRMCA’s membership directory.
6Price is given by sales of concrete divided by tons of concrete sold, where I use data from the materialtrailer to the Census of Manufacturers. I follow Syverson (2004)’s procedure, which removes hot and cold deck
5
��
��
��
�� ����
��
��
�� ����
��
��
��
����
��
��
����
��
��
��
����
��
��
��
����
��
��
��
��
�� �� ��
���� ����
��
����
����
��
��
����
��
��������
����
������ ��
��
����
��
��
������ ������
���� ����
��
��
���� ��
�������� ��
��������
����
������
��
����������
����
��
����
����
����
������
��
��
������
��
��
������ ��
��
��
��
��
��
��
��
��
��
��
��
����
������
��
�� ��
��
��
������ ����
����
��
��������
��
��
�� ����
��
��
��
��
��
����
����
��
��
��
��
��
��
����
��
����
��
��
��
��
��
��
����
��
��
��
��������
����
����
����
�� ��
��
�� ����
���� ���� ��
��
��
��
�� ��
����
���� ��
��
��
����
����
����
������
��
����
��
�� ����
����
��
��
�� ��
����
����
����
��
��
��
����
��
��
��
��
��
��
��
����
����
��
����
������
����
��
��
����
��
����
����
����
������
��
����
��
�� ����
����
��
����
�� ��
�� ������
��
������
����
��
����
������
����������������� ������������������ ��� ��� ��� ������� �����
��� � ��� ��� �����
Source: Zip Business Patterns publicly available dataset at http://www.census.gov/epcd/www/zbp_base.html for NAICS Code 327300.
Figure 1: Dispersion of Ready-Mix Concrete Plant Locations in the Midwest.
6
Number of Concrete Plants Number of Counties/Years Percent
0 22,502 30%
1 23,276 31%
2 12,688 17%
3 6,373 9%
4 3,256 4%
5 1,966 3%
6 1,172 2%
More than 6 3,205 4%
Total 74,438
Table 1: Most counties in the United States are served by fewer than six ready-mix concreteplants.
competitors have little additional impact.7
2.2 Concrete Demand
Most concrete is purchased for building, so I measure demand with employment in the con-
struction sector. Demand is inelastic because it is a small part of construction costs, as these
do not exceed 10 % of material costs for any subsector in construction. So it is implausi-
ble that the ready-mix market substantially affects the volume of construction activity. As
such, changes in construction activity that affect the ready-mix concrete industry’s market
structure are the main source of exogenous variation.
There are large fluctuations in concrete purchases. The autocorrelation of log county
construction employment is 85% for one year, 65% for five years, and 21% for 20 years.
This low autocorrelation of construction activity indicates significant year-to-year variation
in demand.8
imputes by dropping all price pairs that are exactly the same. Section G of the Online Appendix discussesthe construction of price statistics in more detail.
7Caution should be exercised when interpreting these price regressions, as the number of firms couldbe positively correlated with a market that has unusually high prices, as discussed in Manuszak and Moul(2008), so the results in Figure 2 most likely underestimate the price-competition relationship.
8However, the demand process has more long-term correlation than an AR process would predict, as anAR(1) process would predict a four percent 20 year autocorrelation, given an 85 percent one-year autocor-
7
Price and Competition
40
40.5
41
41.5
42
0 1 2 3 4 5 6 7 8
Number of Competitors
Med
ian
Pric
e of
a c
ubic
yar
d of
con
cret
e in
19
63 d
olla
rs
Note: Bars represent 95% confidence interval on median price.
Figure 2: Price declines with the addition of the first competitors but drops by very littlethereafter.
The spatial autocorrelation of demand is also negligible. Only 2.1% of the variation
in log construction employment in a county is accounted for by changes in log construction
employment in counties that border it.9 At the county level, we can think of demand evolving
autonomously and, thus, focus on policies that interfere with county-level demand patterns,
rather than state or national patterns.
2.2.1 Government
Governments purchase half of all U.S. concrete, primarily for road construction.10 These
purchases fluctuate, due to the discretionary nature of highway spending in state and federal
relation. To capture long-run differences in market size, I estimate the process for demand separately fordifferent markets, as discussed in section 5.3.
9Moreover, any aggregate component of construction employment would show up as spatial autocorrela-tion of changes in construction activity.
10According to Kosmatka, Kerkhoff, and Panarese (2002) p.9, government accounts for 48 percent ofcement consumption, with road construction alone responsible for 32 percent of the total.
8
budgets. Government demand is a major source of uncertainty for ready-mix producers.
2.3 Sunk Costs
Opening a concrete plant is an expensive investment. In interviews, managers of ready-
mix plants estimate the cost of a new plant at between three and four million dollars, and
continuing plants in 1997 had on average two million dollars in capital assets. Yet, there
are few expenses involved in shutting down a ready-mix plant. Trucks can be sold on a
competitive, used-vehicle market, and land can be sold for other uses. The plant itself is
a total loss. At best, it can be resold for scrap metal, but many ready-mix plants are left
on-site because the cost of dismantling them outweighs their resale value. 11
3 Data
I use data on ready-mix concrete plants provided by the Center for Economics Studies at
the United States Census Bureau. My primary source is the Longitudinal Business Database
(henceforth, LBD) compiled from data used by the Internal Revenue Service to maintain
business tax records. The LBD covers all private employers on a yearly basis from 1976 to
1999 and has information about employment and salary, along with sectoral coding and firm
identification, but does not record sales, materials, or capital.
Production of ready-mix concrete for delivery predominantly takes place at establish-
ments in the ready-mix sector (NAICS 327300 and SIC 3273), so I choose the establishments
in this sector.12
11I provide evidence of sunk costs in the ready-mix industry, including factors difficult to quantify, such aslong-term relationships with clients and creditors. These intangible assets may account for a large fractionof sunk costs. For instance, ready-mix operators sell about half of their production with a six month graceperiod for repayment. These accounts receivable have a value equivalent to half of a plant’s physical capitalassets. They also function as a sunk cost, as it is more difficult to collect these accounts if the firm cannotpunish non-payment by cutting off future deliveries of concrete.
12Plants occasionally switch in and out of the ready-mix concrete sector. I select all plants that have
9
To construct longitudinal linkages across plants over time, I adapt the Longitudinal Busi-
ness Database Number (henceforth, LBDNUM), developed by Jarmin and Miranda (2002).
This identifier is constructed from Census ID, employer ID, and name and address matches
of all plants in the LBD. I use Jarmin and Miranda (2002)’s plant birth and death flags
to measure entry and exit.13 Each year, about 40 plants (or about 1.6 percent of plants)
are temporarily shut down. I do not treat temporary shutdown as exit, since the cost of
reactivating a plant is smaller than building one from scratch.
I complement the LBD with data from the Census of Manufacturers (henceforth, CMF)
and Annual Survey of Manufacturers (henceforth, ASM), which contain more detailed infor-
mation on plants, such as inputs, outputs, and assets.14 To obtain data on construction, I
select all establishments from the LBD in the construction sector (SIC 15-16-17) and aggre-
gate them to the county level.
The plants in my sample produce 94 percent of the ready-mix concrete shipped in the
manufacturing sector. Moreover, for these plants, ready-mix concrete is 95 percent of their
output.
belonged to the ready-mix sector at some point in their lives but disregard plants that switch into theconcrete sector for only a small fraction of their lives, since these transient concrete plants are typicallymiscoded and manufacture products such as cement or concrete pipe. Specifically, I exclude from my sampleplants that produce concrete less than 50% of the time.
13Jarmin and Miranda (2002) identify entry and exit based on the presence of a plant in the I.R.S.’s taxrecords. They take special care to flag cases where plants simply change owners or names by matching theaddresses of plants across time. If a plant changes ownership, I do not treat this as an exit event, since thecost of changing the management at a plant should be much lower than the cost of building a plant fromscratch.
14Unfortunately, the ASM is only sent to about one-third of plants in the ready-mix concrete sector, whilethe CMF is available only every five years and excludes all plants with fewer than five employees (i.e., aboutone-quarter of concrete plants). Since the CMF and ASM have serious issues with missing data, it is difficultto use them alone for longitudinal market-level studies. This is not true of the LBD, which includes theentire population of U.S. plants.
10
3.1 Panel
Over the sample period of 1976 to 1999, there were about 350 plant births and 350 plant
deaths each year, compared to 5,000 continuers. Both turnover rates and the total number
of plants were stable over the period.
The average ready-mix concrete plant employed 26 workers and sold about $ 3.4 million
of concrete in 1997. About half of all sales are accounted for by material costs, while the rest
is value added. However, these averages mask substantial differences between plants. Most
notably, the distribution of plant size is heavily skewed, with few large plants and many small
ones. For instance, five percent of plants have one employee, and less than five percent of
plants have more than 82 employees. Continuing plants are twice as large as either entrants
(i.e., births) or exitors (i.e., deaths), measured by capitalization, salaries, or shipments.
Plant size is a crucial difference between plants, as bigger plants ship more concrete and
are far less likely to exit. I use employment to measure size at a plant. Employment is a
better measure of size than capital stock, since it is available for all plants in the sample,
while capital stock is available for less than 20% of plant-years. Moreover, the number of
employees is more autocorrelated than capital assets (91 % versus 74 % for capital), and is a
better predictor of both future production (with a correlation of 92 % with total shipments
versus 43 % for capita), and the likelihood of exit.1516
I call a plant small if it has fewer than eight employees, medium if it has between eight
15I use employment instead of capital stock, since employment is measured for all plants in the data (itis derived from IRS tax returns in the LBD), while capital is available for all plants in a market for only asmall number of markets (as is discussed in Collard-Wexler (2009), which uses multiple imputation to fill inmissing capital stock). In practice, given the coarseness of my employment bins, classifying a firm based oncapital or employment does not matter very much.
16Ready-mix concrete has been studied extensively by Syverson (2004), who provides evidence of productiv-ity dispersion across plants. In another paper, Collard-Wexler (2009), I discuss the dynamics of productivitydispersion. Incorporating productivity differences into the model leads to a great number of data challenges,as the data needed to construct productivity is frequently missing or imputed. Moreover, incorporatingproductivity into a dynamic model leads to a focus on variability in plant-level productivity, since thesevariations in productivity dwarf variations in demand.
11
Current Size Size Next YearOut Small∗ Medium∗∗ Large∗∗∗ Total
Out 98.5% 1.2% 0.2% 0.1% 15,134Small 8.0% 82.4% 8.1% 1.5% 1,135Small, Medium in Past 7.9% 73.6% 17.6% 1.0% 417Small, Large in Past 11.7% 65.8% 16.4% 6.1% 140Medium 3.2% 20.1% 68.6% 8.1% 686Medium, Large in Past 3.2% 11.0% 64.4% 21.3% 307Large 2.7% 4.1% 11.1% 82.1% 913
* Small: Less than 8 Employees, ** Medium: 8 to 17 Employees,*** Large: More than 17 Employees
Table 2: Average Yearly Plant Transition Probabilities
and 17 employees, and large if there are more than 17 employees.17 I also keep track of the
largest size a plant attains, since a plant that was previously large may have assets that make
it easier for it to ramp up in the future. Table 2 shows the probability that plants will change
size, enter, or exit. In the sample of counties that excludes large markets, 47% of plants are
small, 28% are medium, and 25% are big. The history of a plant’s size matters. Large plants
exit at a rate of 2.6% - one-third the rate of small plants (8.0%) - and plants that were large
in the past are more likely to expand in the future. For instance, a medium-size plant that
was large in the past has a 21% probability of becoming large next year, versus only an 8%
probability for a plant that has never been large.18
I aggregate plant data by county to form market-level data. Since counties in the United
States vary greatly in size, I have taken care to exclude counties in states, such as Ari-
zona, that have unusually spacious counties and a small number of heavily populated urban
counties.19
17I choose cutoffs of eight and 18 employees because these correspond to the 33rd and 66th quantiles ofthe empirical distribution of employment.
18I do not keep track of past size if a plant is larger today than it was in the past, since it is the largestsize of previous employment that determines if a firm has the equipment and land necessary to ramp up inthe future. As well, if I kept track of past size, regardless of current size, this would increase the number ofplant-level states from seven to 10, raising the size of the state space for the entire industry by a factor ofabout 14.
19Specifically, I exclude counties with more than 20 ready-mix concrete plants, which are all urban areas.The County Business Patterns reports that there were 20 of these counties in 2007.
12
Table 3 presents summary statistics of the market-level data. On average, there are 1.86
plants per market. Moreover, there is a wide range of construction employment, from 11
employees (5th percentile) to 6,800 employees (95th percentile).20
Observations Mean Standard 5th 95thDeviation Percentile Percentile
Concrete Plant DataConcrete Plants 74435 1.86 3.24 0 6Employment 74435 27.24 79.03 0 110Payroll (in 000’s) 74435 4238 74396 0 3600Total Value of Shipment (in 000’s) 24677 3181 12010 0 14000Value Added (in 000’s) 24677 1408 5289 0 6500Total Assets Ending (in 000’s) 24677 1090 14134 0 4700
Construction Establishment DataEmployment 69911 1495 5390 11 6800Payroll (in 000’s) 69911 37135 163546 110 160000
County Area (in square miles) 72269 1147 3891 210 3200
Table 3: Summary Statistics for County-Aggregated Data
4 Model
I adapt the theoretical framework for dynamic oligopoly developed by Ericson and Pakes
(1995) to analyze entry, exit, and investment decisions in the ready-mix concrete industry. In
each market, there are i = 1, · · · , N firms, which are either potential entrants or incumbents.
A firm i can be described by a firm-specific state sti ∈ Si. The typical ready-mix firm owns
a single plant, and I assume that each firm owns a single ready-mix concrete plant, making
plant and firm interchangeable.21,22 Firms also react to market-level demand M t and thus,
20 Yet, the range of the surface area of counties, in square miles, falls between 210 and 3,200 - a 10-to-onedifference - versus a 500-to-one difference for construction employment.
21Indeed, Syverson (2004) reports that 3,749 firms controlled the 5,319 ready-mix plants operating in 1987.22Due to antitrust policy in the United States, ready-mix concrete firms historically were prevented from
merging with upstream cement producers. In most other countries, ready-mix concrete plants are verticallyintegrated with cement producers.
13
the market level state st is the composition of the states for each firm and the aggregate
state M t:
st = {st1, st2, · · · , stN ,M t}
I distinguish between two components of the state sti: xti, which is common knowledge to
all firms in the market, and εti, which is an i.i.d. private information component.23 Denote
by xt = {xt1, xt2, · · · , xtN ,M t} and εt = {εt1, εt2, · · · , εtN}, the market-level common knowledge
and private information state respectively.
The difficulty in dynamic games is in computing an equilibrium for counterfactuals. More
precisely, for this application the main burden is keeping the entire state space in memory. I
choose a maximum of 10 plants per market, since this allows me to pick up most counties in
the United States (where the 95th percentile of the number of plants in a county in Table 3
is six), and keeps the size of the state space manageable. A county with more than 10 active
plants at some point in its history is dropped from the sample, since the model does not
allow firms to envisage an environment with more than nine competitors.24
Firm i can be described by a firm specific state sti ∈ Si:
sti = { xti︸︷︷︸(Plant Size,Past Plant Size)
, εti︸︷︷︸i.i.d. shock
} (1)
a vector of i.i.d. unobserved shocks εti. The firm’s observed state xti is whether a firm is big,
medium, or small, as well as the firm’s largest previous size, as described in Table 2.
In each period t, potential entrants choose whether or not to enter a market, and incum-
23If instead εti was serially correlated, then a firm might find it optimal to condition its strategy on pastactions taken by other firms in the market. This would substantially increase the size of the state space. Forinstance, if a firm conditioned its strategy on the history of the market for even a single year, the state spacewould be more than 1.9 quintillion.
24To allay the potential for selection bias that this procedure entails, counties with more than 10 000construction employees at any point between 1976 and 1999 are also dropped. This excludes 15% of marketsand 35% of plants from the analysis.
14
bents can choose to exit the market. Conditional on being in the market, firms pick their
common knowledge state xti in the next period. Thus, the firm’s action ati is the choice of
being out of the market, i.e. xt+1i = ∅, or their state tomorrow xt+1
i (small, medium, large).
Demand evolves following a first-order Markov Process with transition probabilities given
by D(M t+1|M t).
I assume the private information vector εti enters into the profit function as an additive
logit shock to the value of each action ati. Payoffs are given by:
r(xt+1) + τ(xt+1i = ati, x
ti) + εtia (2)
where r(·) denote the rewards from operating in the market, and τ(·) are transition costs
- i.e., the costs of moving from one state to another. The results section of this paper is
primarily concerned with estimating these reward and transition functions.
The game’s timing is:
1. Firms privately observe εti and publicly observe xt.
2. Firms simultaneously choose actions ati.
3. Demand M t evolves to its new level M t+1. Firm level states evolve to xt+1i .
4. Payoffs r(xt+1) + τ(ati, xti) + εtia are realized.
I define the firm’s ex-ante (i.e., before observing εti) value as:
V (xt) = Eεti
(maxati
Ext+1−i
[r(xt+1−i , x
t+1i ) + τ(xt+1
i = ati, xti) + εtia + βV (xt+1)
])(3)
and firms pick the action that maximizes the net present value of rewards:
a∗ti = argmaxatiExt+1−i
[r(xt+1−i , x
t+1i ) + τ(xt+1
i = ati, xti) + εtia + βV (xt+1)
](4)
15
Doraszelski and Satterthwaite (2010) show that if εtia is an additive, action-specific shock
that has full support, then there will exist pure strategy Nash equilibria for this game - i.e.,
policies a∗(xt, εti) such that a unilateral, one-shot deviation to strategy ai(xt, εti) does not
lead to a higher net present value of rewards, conditional on all other players using strategies
a∗−i(·).25 Last, I introduce some additional notation. To work out the firm’s strategies, I
compute the ex-ante choice specific value function W (ati, xt) - i.e., the net present value of
payoffs conditional on taking action ati before εti is observed, defined as:26
W (ati, xt) = Ext+1|ati
[r(xt+1) + τ(ati, x
ti) + βV (xt+1)
]= Ext+1|ati
[r(xt+1) + τ(ati, x
ti) + βEεt+1 max
at+1i
(W (at+1
i , xt+1) + εt+1ai
)] (5)
Given the choice-specific value function, it is easy to reckon the firm’s conditional choice
probability (henceforth CCP) Ψ[ati|xt], i.e., the probability that a firm will play action ati in
an observable state xt - before observing εti - using the logit formula:
Ψ[ati|xt] =exp (W (ati, x
t))∑j∈Ai
exp (W (j, xt))(6)
25Proposition 2 in Doraszelski and Satterthwaite (2010) describes conditions under which the Ericson andPakes (1995) model has a pure strategy equilibrium, essentially pointing out that exit and entry costs needto have full support shocks to ensure the existence of a pure strategy equilibrium. The game I describe hasfull support shocks to the value of entering and exiting, as well as to the value of taking any action.
26Remember that V and W are linked together by:
V (xt) = Eεati
(maxati
W (ati, xt) + εati
)
16
5 Parametrization and Conditional Choice Probability
(CCP) Estimation
In this section, I will first present the parametrized profit function that I estimate. Second,
I show the Indirect Inference Conditional Choice Probability estimator (henceforth IICCP)
used to recover these parameters. Third, I discuss details of the estimation of conditional
choice probabilities.
5.1 Profit Function
I use a simple Bresnahan and Reiss (1991) style reduced-form for the reward function, en-
dowed with parameters θ that I estimate:
r(ati, xt|θ) =
∑α∈{Big,Medium,Small}
1(ati = α)
×
θα1︸︷︷︸Fixed Cost
+ θα2Mt+1︸ ︷︷ ︸
Demand Shifter
+ θα3 g(∑−i
xt+1−i 6= out)︸ ︷︷ ︸
Competition Parameters
(7)
where g(·) is a non-parametric function of the number of competitors. This reward function
is linear in parameters that I exploit during estimation.27
Transition costs are:
τ(ati, xti|θ) = θl,m4
∑l>0,m 6=l
1(ati = l, xti = m) (8)
so a firm pays a transition cost to change its state. However, I assume that a firm does not
27This “reduced-form” profit function is an approximation to the profits earned by competitors. For es-timation purposes I need to assume that the specification error in r is orthogonal to the state variablesx.
17
pay any exit costs.28
Sections 5.2 and 5.3 discuss the estimation of the parameters θ of the profit and transition
cost function. The reader can skip to section 6 for estimates of these parameters.
5.2 Indirect Inference CCP Algorithm
Applying the Ericson and Pakes (1995) framework to data has proven difficult, due to the
complexity of computing a solution to the dynamic game and multiple equilibria.29 For single-
agent problems, Hotz and Miller (1993) and Hotz, Miller, Sanders, and Smith (1994) bypass
the computation of optimal policies by estimating policies directly from agents’ choices. This
idea has been adapted to strategic settings by several recent papers in Industrial Organization
- most prominently, Bajari, Benkard, and Levin (2007), Pakes, Berry, and Ostrovsky (2007),
Pesendorfer and Schmidt-Dengler (2008), Ryan (2012), and Dunne, Klimek, Roberts, and
Xu (2006).
I estimate the model by matching the optimal choice probabilities Ψ(ai|x, θ) to the data.
The natural way to do this would be to compute an equilibrium to the dynamic game’s given
parameters θ. However, doing this for each candidate parameter vector θ is computationally
impractical.
Instead, I have adapted a conditional choice probability estimator that can be applied
to games. My CCP algorithm can handle the very large state space in this problem (over
350 000 states), and I use a Simulated Indirect Inference Criterion approach for estimation
(Keane and Smith Jr (2003),Gourieroux and Montfort (1993) and Gourieroux and Monfort
(1996)).30
28While entry, fixed costs, and exit costs are not strictly collinear, Monte Carlo experiments indicate thatit is quite difficult to jointly identify all three of these costs. Section D of the Online Appendix shows theidentification of this model, as well as some intuition for why it is difficult to separately identify fixed costs,entry costs, and exit cost.
29Even with the high performance Stochastic Algorithm used in this paper, it takes more than an hour tocompute a solution.
30In previous version of this paper, I computed present estimates using an approach in the spirit of
18
Algorithm CCP Indirect Inference Algorithm (CCPII)
1. Replace optimal choice probabilities Ψ with an estimate from the data P .
Estimate the demand transition process D[M t+1|M t].
I assume a single symmetric Markov-Perfect equilibrium played in each observed state
x. Thus, I can recover the empirical analogue to Ψ by looking at the empirical frequency
of actions in different states P = {Pr(ati|xt)}ati,xt . Likewise, I can estimate the demand
transition process (denoted D) using the observed demand transitions in the data. I
discuss the details of the estimation of P and D in section 5.3.
2. Compute the W function up to a vector of parameters θ, conditional on
policies Ψ(ati|xt) = P [ati|xt].
A final rewriting of the W-function is now in order to aid with the estimation of the
model. The rewards and transition costs in equations (7) and (8) on page 17 are linear
in parameters θ, so the profit function can be rewritten as r(ai, x|θ) − τ(ai, xi|θ) =
θ ·~ρ(ai, x), where ~ρ is a function that returns a vector. This implies that the W function
is separable in dynamic parameters as in Bajari, Benkard, and Levin (2007), since
W (ai, x|θ) = E∞∑t=1
βt(r(ati, x
t|θ)− τ(ati, xti|θ))
= θ · E∞∑t=1
βt~ρ(ati, xt) ≡ θ · Γ(ai, x)
(9)
Note that the Γ function only depends on the expected evolution of the state and
actions in the future, rather than on the parameter vector θ:
Γ(ai, x) = E∞∑t=1
βt~ρ(ati, xt) (10)
Aguiregabiria and Mira (2007), which iteratively updates the strategies used by firms. I find that using aniterated technique yields very similar results to those presented in the paper.
19
I compute the Γ function using forward simulation, in which I simulate the evolution of
the state x and action ai by drawing from the choice probabilities Ψ and the demand
transition process D. Since I have replaced these objects by their empirical analogues P
and D, I can perform this forward simulation without solving the model. The forward
simulation is done with a Discrete Action Stochastic Algorithm (henceforth DASA)
that is close to Pakes and McGuire (2001) presented in Online Appendix B.
The optimal choice probabilities Ψ can be rewritten as a function of ΓP and θ (where
I include the subscript P to emphasize that Γ depends on my estimate of CCPs):
Ψ(ai|x,ΓP , θ) =exp
(θ · ΓP (ai, x)
)∑
j∈A exp(θ · ΓP (j, x)
) (11)
3. Simulated Indirect Inference Estimation
I use an Indirect Inference Criterion function to estimate the model.31 The estimator
matches regression coefficients from the data (denoted β) with regression coefficients
from simulated data generated by the model, conditional on a parameter θ (denoted
β(θ)). I use a multinomial linear probability model as an auxiliary model. It is simple
to estimate and is a close analogue to the multinomial dynamic logit model.32
I define the outcome vector from the data as yn and the predicted choice probabilities
31Indirect Inference is less sensitive to error in the Γ function than Maximum Likelihood and, like manyGMM estimators, can be consistent even if there is simulation error in Γ, and this simulation error does notvanish asymptotically. For some intuition, if the exit rate in the data is 1%, but the model predicts an exitprobability of almost 0%, then a maximum likelihood criterion would have an infinite log-likelihood, whilean Indirect Inference criterion would find an error of 1%. I find it easier to minimize this criterion functionversus a criterion of the form ‖yn − yn(θ)‖, which is closer to traditional GMM.
32The auxiliary model does not need to be a consistent estimator and need not have an interpretation ofany sort. Its sole responsibility is to provide rich description of the patterns of a dataset and to be simple toestimate.
20
given by the model yn(θ) for observation n as:
yn =
1(an = small)
1(an = medium)
1(an = big)
yn(θ) =
Ψ(small|xn,Γ, θ)
Ψ(medium|xn,Γ, θ)
Ψ(large|xn,Γ, θ)
(12)
where the outcome vector yn(θ) is the predicted choice probabilities Ψ.33 I run an OLS
regression on yn = Znβ and find the OLS coefficients of the multinomial linear prob-
ability model. Likewise, I run an OLS regression on the predicted choice probabilities
yn to obtain the coefficients for the model β(θ), given parameter θ.
The criterion function minimizes the distance between the regression coefficient in the
data and in the simulated data:
Q(θ) =(β − β(θ))
)′W(β − β(θ)
)(13)
where W is a weighting matrix. I use W = V ar[β]−1, the inverse of the covariance
matrix from the OLS regression. Section D of the Online Appendix shows conditions
under which the estimator is consistent, which is an extension of the consistency of
Indirect Inference estimators.34
33 Theorem 1 in Section D of the Online Appendix proves that using the choice probabilities Ψ as predictedactions gives the same θ’s as drawing action an ∼ Ψ(·|xn,Γ, θ) from the predicted choice probabilities whenone uses an infinite number of simulation draws.
34In a prior version of the paper, I estimated the model by iterating on the conditional choice probabilities -i.e., updating them using parameter estimates θ. To implement this procedure (which requires the assumptionof a single equilibrium for the dynamic game in order to be consistent), I need to add extra steps where:
4) Replace P [ai|x] with Ψ[ai|x, θ] where θ is the current estimate of the parameters in the profit function
and Ψ[ai|x, θ] are computed equilibrium policy functions given θ.5) Repeat steps 2-4 until θ converges.When I iterate on the conditional choice policies, I get results that are very similar to the results obtained
when I do not.
21
5.3 Conditional Choice Probabilities
5.3.1 Detecting Market Unobservables
I assumed that the unobserved state εati is an i.i.d. logit, which rules out persistent market-
level unobservables. This is a problem. Some markets have higher costs than others, due to,
for instance, the presence of unionized workers in Illinois but not in Alabama. In addition,
some markets have higher demand for concrete that is not captured by employment in the
construction sector - for instance, because asphalt, but not concrete, melts on roads in Texas
but not in Maine. 35
To detect market unobservables, Table 4 runs binary logit regressions of a plant’s decision
to be active in the market (i.e. have a plant) on construction employment, the number of
competitors, plant size in the prior year, and largest ever plant size. Column I presents the
base estimates, Column II includes market-fixed effects via a conditional logit, and Column
III has state- and year-fixed effects. Column IV includes indicators for market categories µ,
which I henceforth refer to as market-category effects. These categories are constructed by
rounding the average number of plants in a county to the nearest integer. Finally, Columns
V, VI, VI, and VII show alternative category controls based on the lagged average number
of firms, the average number of firms before 1983 (on data from 1984 to 1999), average log
construction employment, and average total shipments of concrete, again grouped into four
categories.36
The effects of the second, third, and additional competitors are close to zero in Column
I; they turn negative when I include market-fixed or market-category effects in Columns
II and IV. If the market-level shock is ignored, then the number of competitors will be
positively correlated with market unobservables. This leads to upward bias in the competition
35There are numerous differences between markets, such as their road network, intensity of use of concretein construction, density, area served, and input costs for cement and gravel. Thus construction employmentalone cannot possibly capture all components of a market’s profitability.
36A more thorough discussion of market-category controls can be found in Online Appendix C1.
22
Dependent Variable: Activity I II (FE†) III IV (µ) V VI VII VIII
Log County 0.17*** 0.03 0.19*** 0.02* 0.03*** 0.20*** 0.09*** 0.07***Construction Employment (0.01) (0.02) (0.01) (0.01) (0.01) (0.01) (0.02) (0.01)First Competitor -1.40*** -1.26*** -0.85*** -1.07*** -1.01*** -0.69*** -0.73*** -0.85***
(0.05) (0.04) (0.04) (0.04) (0.04) (0.01) (0.04) (0.04)Second Competitor 0.00 -0.54*** -0.03 -0.48*** -0.47*** -0.03 0.04 -0.03
(0.04) (0.04) (0.04) (0.04) (0.04) (0.04) (0.04) (0.04)Third Competitor 0.03 -0.33*** 0.04 -0.32*** -0.32*** 0.00 0.07 0.03
(0.04) (0.04) (0.05) (0.05) (0.05) (0.06) (0.05) (0.05)Log Competitors above 4 0.02 -0.13*** 0.09** -0.06* -0.10* 0.10** 0.10** 0.09**
(0.03) (0.03) (0.03) (0.03) (0.03) (0.05) (0.03) (0.03)
Small 6.89*** 6.50*** 6.92*** 6.73*** 6.75*** 7.06*** 6.89*** 6.91***(0.04) (0.03) (0.04) (0.04) (0.04) (0.05) (0.04) (0.04)
Small, 6.85*** 6.36*** 6.82*** 6.60*** 6.61*** 6.92*** 6.85*** 6.83***Medium in Past (0.05) (0.04) (0.05) (0.05) (0.05) (0.06) (0.05) (0.05)Small, 6.41*** 5.90*** 6.34*** 6.19*** 6.19*** 6.42*** 6.40*** 6.36***Large in Past (0.07) (0.06) (0.07) (0.07) (0.07) (0.08) (0.07) (0.07)Medium 7.72*** 7.34*** 7.66*** 7.54*** 7.56*** 7.95*** 7.72*** 7.66***
(0.06) (0.05) (0.06) (0.06) (0.06) (0.08) (0.06) (0.06)Medium, 7.72*** 7.26*** 7.60*** 7.47*** 7.47*** 7.82*** 7.71*** 7.63***Large in Past (0.08) (0.08) (0.08) (0.08) (0.08) (0.10) (0.08) (0.05)Large 7.85*** 7.48*** 7.71*** 7.63*** 7.64*** 7.91*** 7.83*** 7.73***
(0.06) (0.05) (0.06) (0.06) (0.06) (0.07) (0.06) (0.06)
Market Fixed Effects XState Fixed Effects XYear Fixed Effects XMarketClassification Variable�Average Number of Plants XLagged Average Plants XBefore 1983 Average Plants XConstruction Employment XTotal Shipments of Concrete X
Observations 409850 409850 409850 260170 409850 409850 409850 409850Markets 2029 2029 2029 2029 2029 2029 2029 2029Log-Likelihood -37541 -32759 -37429 -36713 -36695 -22230 -37524 -37384χ2 49545 301019 51517 52775 51795 34019 50302 51538
Note: Standard Errors are Clustered by Market. † Market-fixed effects are implemented via a conditional
logit. � Average number of plants (µ) is the mean number of plants in a market, rounded to the nearest integer.
Lagged Average Plants is the mean number of plants in the market for years preceding t, rounded to the
nearest integer. Before 1983 Average Plants is the mean number of plants in a market before 1983, rounded
to the nearest integer. Only years from 1983 to 1999 are used in the regression for this market classification
variable in Column V. Construction Employment Classification and Total Shipments of Concrete use the
mean of these variables in a market to classify markets into four categories.
Table 4: Binary Logit Regressions of the decision to have an active plant with market-fixedeffects and market-category effects.
23
coefficient.
The effect of past plant size has a substantial effect of the probability of activity today,
which is to be expected, given the sunk costs of opening a ready-mix concrete plant. However,
this effect of past size is smaller in Columns II and IV than in Column I. Past plant size is
also related to market size - both observed and unobserved. Thus past plant size also proxies
for serial correlation in unobserved market demand and is biased upwards.
The effect of log country construction employment falls when market-category or market-
fixed effects are added. Market effects wash out a large part of the correlation between
demand and the number of firms in a market, since much of this comes from cross-sectional
variation. As discussed in the context of production function estimation by Griliches and
Hausmann (1986), the remaining time-series variation in demand is more likely to suffer
from measurement error, which attenuates the demand coefficient. Much of the treatment of
market-category effect to follow is a “hack” to navigate the twin issues of upwardly biased
competition coefficients and attenuated demand coefficients.37
Adding year- and state-fixed effects (Column III) does not substantially change the co-
efficients from Column I (no effects). Heterogeneity across markets is the issue, rather than
year-to-year shocks. Moreover, while there are large differences between states in their use
of ready-mix concrete, these do not capture much of the differences between markets.
5.3.2 Market Categories
While estimating market-level policy functions is straightforward, it runs into serious data
constraints, since I cannot identify parameters from the cross-section. To render the market-
37This problem is similar to the use of fixed effects in a production function regression discussed in Grilichesand Hausmann (1986), where fixed effects eliminate the most important source of variation in capital stock,thereby leading to a downward bias on the capital coefficient. These two biases are a serious problem: havingno market-level controls leads to a market where the number of firms sloshes around, since competitioneffects are too small to pin down the number of firms, while a model with market-fixed effects predicts aresponse to demand that is too small and too few changes in the number of plants.
24
fixed effects tractable, I collapse market effects into market-category effects µ. This classifi-
cation scheme is based on an endogenous variable, but the estimates in Column V that use
the lagged number of firms, a variable that is not endogenous, are indistinguishable from
the market-category estimates in Column IV.38 Likewise, Section C2 of the Online Appendix
shows estimates based on categories constructed from estimated market-fixed effects. These
types of categories yield similar estimates as those in Column III using market categories µ.
However, other plausible market-category controls do not work. These include the number
of firms in a pre-period (Column VI), average log construction employment (Column VII),
total shipments of concrete (Column VII), and the square mileage of a county and its density.
They are similar to the estimates without market controls in Column I.
5.3.3 Estimating CCP’s
Since the εtai ’s are logit draws, I estimate the conditional choice probabilities with a multi-
nomial logit of a firm’s choice of its size next year (ati), presented in Table 5. Column II has
market-category effects µ (henceforth P µ), but Column I does not (henceforth P ).39
These multinomial logits illustrate the identification in the model. Firms are more likely
to exit, or less likely to enter, if there are more competitors or higher demand. Second, past
plant size explains the a firm’s current choice of size and activity, indicating the role of sunk
costs. Finally, for this model the ε shocks are not inconsequential, as these generate entry
and exit not connected to observable shifters of profits; demand and competition.
As seen in the findings from Table 4, introducing market-category effects leads to sig-
nificantly more negative effects of competitors. In particular, the effect of more than one
38However, this classification scheme, which is based on the lagged number of firms, is harder to fit intothe model, since a market can switch categories, which makes the market-category a state variable.
39Due to limited data, rather than estimating coefficients on the logit βµ,0ai +βµ,Xai X that all vary by marketcategory, I assume that the market effects are just additive constants - i.e., βµ,0ai + βXaiX. The main issue isthat it is difficult to estimate the effect of, say, the third competitor in a market that has, on average, onefirm in it - hence in market category µ = 1, as we rarely see three firms in this type of market.
25
competitor is positive without market-category effects. Positive effects of competition have a
toxic effect on both estimation and counterfactuals, since simulating the model forward with
positive spillovers between firms makes the market tip from no firms to being completely
filled up with firms.
Finally, both the no-effect and category effects estimates show that the effect of demand
is much higher for large plants than small ones, as the effect of log construction employment
is 0.13 for small plants, versus 0.29 and 0.51 for medium and large plants respectively. This
happens because larger markets have bigger plants, and I return to this issue in section 6.40
5.3.4 Estimating Demand Transitions
The demand transition matrix D is estimated by market category µ using a bin estimator
Dµ[i|j] =∑
(l,t) 1(Mt+1l ∈Bi,M
tl ∈Bj)∑
(l,t) 1(Mtl ∈Bj)
with 10 bins (that differ by category); the demand level within
a bin is set to the mean demand level. Note that construction employment varies considerably
across market categories - e.g., from 285 to 4,200 employees from market category 1 to 4.
Finally, the Γµ function is computed by market category µ using CCP’s P µ and demand
transition process Dµ.41
6 Results
I fix the discount factor to 5% per year. The covariates of the multinomial linear probability
model (denoted zn) used to estimate the β coefficients and used in an auxiliary model in the
40It is important for the CCP’s to be able to replicate firms’ expectations over the evolution of the ready-mix concrete market. Section 6.1 discusses this issue.
41I can go one step further and include market categories in a firm’s profit function, which allows me toestimate a profit function rµ(ati, x
t|θ) where rewards are additively separable in the market-category levelcomponent:
rµ(ati, xt|θ) = r(ati, x
t|θ) + ξµa + εati
and have a market/action effect ξµai . I have also estimated this model with market category fixed effect. Ifind that estimating market category profit shifters yields inferior fit as compared to the procedure used inthe paper.
26
Dependent Independent I II (Market Category)Variable Variable Coeff. S.E. Coef. S.E.
Small Small 6.59 (0.03) 6.42 (0.03)in t+ 1 Small Medium in Past 6.45 (0.04) 6.18 (0.04)
Small Large in Past 5.94 (0.06) 5.72 (0.06)Medium 6.02 (0.05) 5.81 (0.05)Medium Large in Past 5.41 (0.08) 5.16 (0.08)Large 4.58 (0.06) 4.37 (0.06)Log County Employment 0.13 (0.01) -0.06 (0.01)First Competitor -1.42 (0.04) -1.71 (0.04)Second Competitor 0.10 (0.03) -0.46 (0.03)Third Competitor 0.16 (0.04) -0.26 (0.04)Log of Competitors above 3 0.11 (0.03) -0.04 (0.03)Market Category µ XConstant -3.94 (0.06) -3.17 (0.06)
Medium Small 6.25 (0.05) 6.08 (0.05)in t+ 1 Small Medium in Past 6.96 (0.06) 6.70 (0.06)
Small Large in Past 6.43 (0.08) 6.22 (0.08)Medium 9.16 (0.06) 8.96 (0.06)Medium Large in Past 9.08 (0.08) 8.83 (0.08)Large 7.44 (0.07) 7.23 (0.07)Log County Employment 0.29 (0.01) 0.12 (0.01)First Competitor -1.54 (0.05) -1.87 (0.05)Second Competitor 0.00 (0.04) -0.53 (0.04)Third Competitor 0.06 (0.05) -0.32 (0.05)Log of Competitors above 3 0.02 (0.03) -0.11 (0.03)Market Category µ XConstant -6.72 (0.08) -5.99 (0.09)
Large Small 5.04 (0.08) 4.88 (0.08)in t+ 1 Small Medium in Past 4.53 (0.13) 4.28 (0.13)
Small Large in Past 5.78 (0.11) 5.58 (0.11)Medium 7.46 (0.08) 7.27 (0.08)Medium Large in Past 8.37 (0.09) 8.13 (0.09)Large 9.76 (0.07) 9.56 (0.07)Log County Employment 0.52 (0.01) 0.34 (0.02)First Competitor -1.61 (0.05) -1.94 (0.06)Second Competitor -0.03 (0.05) -0.58 (0.05)Third Competitor -0.02 (0.06) -0.42 (0.06)Log of Competitors above 3 -0.04 (0.04) -0.17 (0.04)Market Category µ XConstant -8.58 (0.11) -7.83 (0.11)
Observations 409850 409850Log-Likelihood 84855 83814Likelihood Ratio 400760 402841
Table 5: Multinomial Logit on the choice to be large, medium or small.
27
CCPII, are indicators for the firm’s current state, the number of competitors in a market,
and the log of construction employement in the county. These coefficients vary by market
category µ and by action chosen ai. Thus the model matches moments conditioned on market
category µ.
Table 6 presents estimates of the dynamic model. In line with interviews with producers
in Illinois, I calibrate the entry costs for a medium sized plant to 2 million dollars. This
allows me to convert parameters in variance units into dollars. In order to make sense of the
magnitudes of these estimates, note that average sales are 3.4 million dollars per year. The
variance of ε is estimated to 133 000 dollars per year, or about 4% of sales, which is below
year-to-year changes in profits due to changes in productivity.42,43 The magnitude of ε is
important, since these i.i.d. shocks generate both turnover and changes in plant size that are
unrelated to changes in demand.
The fixed costs of operating a plant are about $244,000 for a medium-size plant, slightly
less for a small plant, and slightly more for a large plant. Doubling the number of construction
workers in a county increases profits by $6,000 for a small plant versus $11,000 and $14,000
for a medium- and large-size plant, respectively. This reflects the fact that bigger markets
have both more plants and larger plants.
Indeed, Figure 3, which plots local polynomial regressions of the log of construction
employment in the county against both the average size of a plant (measured by payroll)
and the number of plants in a county, shows that plant size is increasing with market size.
For instance, a county with 150 construction sector employees has an average payroll per
concrete plant of $400,000, while a county with 1,010 construction employees has an average
42Collard-Wexler (2009) estimates a similar model that allows for productivity differences between pro-ducers and finds large differences in per-period profits due to differences in productivity.
43I could have used revenues of ready-mix plants to convert my estimates into dollars. When I computeplant-level variable profits as the difference between plant-level revenues and plant-level costs, I obtainimplausibly high rates of return on capital (on the order of 30%). Thus I choose not to use revenue data inthe estimation of the model.
28
(All estimates in thousands of dollars)Coef. S.E.∗
Fixed Cost Small -139 (6)Medium -244 (10)Large -285 (6)
Log Construction Small 20 (1)Employment Medium 35 (2)
Large 45 (1)1st Competitor Small -48 (4)
Medium -58 (5)Large -63 (6)
Log Competitors Small -17 (3)(above 1) Medium -44 (4)
Large -48 (3)Transition CostsOut →Small -1002 (11)Out →Medium † -2000 (107)Out →Large -1771 (53)Small →Medium -332 (7)Small, Past Medium →Medium -772 (32)Small, Past Large →Medium -325 (8)Small →Large -1809 (73)Small, Past Medium →Large -608 (19)Small, Past Large →Large -343 (16)Medium →Small -107 (6)Medium, Past Large →Small -314 (6)Medium →Large 101 (14)Medium, Past Large →Large -43 (7)Large →Small -254 (7)Large →Medium -403 (6)
Standard Deviation of Shock 133
†: The entry costs of a medium sized plant are calibrated to 2 million dollars. ∗: StandardErrors are computed using 100 block bootstraps replications, where I reestimate the demandtransition process Dµ and the conditional choice probabilities P µ, then minimize the criterionfunction Q to find θ. I block bootstrap by market, resampling a market’s history from 1976to 1999, so the computed standard errors account for serial correlation within a market.
Table 6: Estimates for the Dynamic Model of Entry, Exit and Investment.
29
050
010
0015
00Pa
yrol
l Per
Est
ablis
hmen
t in
Thou
sand
s of
Dol
lars
4 6 8 10Log of Country Construction Employment
05
1015
20
Num
ber o
f Est
ablis
hmen
ts
0 2 4 6 8 10Log of Country Construction Employment
Source: County Business Patterns 2007, Construction: NAICS 23. Ready-Mix Concrete NAICS 327320
County Size and the number and size of ready-mix concrete plants
Note: Local Polynomial Regressions with shaded areas representing the 95% confidence interval.
Figure 3: Larger Markets have both more plants and larger plants.
that is closer to $600,000. This effect is not specific to the ready-mix concrete industry, as
Campbell and Hopenhayn (2005) have also documented this link for retail trade. Moreover,
this implies that any change in the size of a market will alter the industry’s size distribution.
Second, there is a linear relationship between the size of establishments and the log of
construction employment for the small markets examined in this paper. Regressing the log
of the number of concrete plants on log construction employment, I find a coefficient of 0.69,
so a 1% increase in construction employment increases the number of firms by less than 1%.
This implies that the number of plants in a county is a concave function of construction
employment. The concavity of the number of plants per market as a function of construction
employment will turn out to have important implications when I run counterfactuals that
reduce demand volatility.
The number of competitors in a county has a large effect on profits. The first competitor
30
reduces profits by $58,000 for a medium-size plant, and doubling the number of competi-
tors (beyond the first competitor) reduces profits by $44,000 per year. The first competitor
has a larger impact than subsequent competitors, which echoes the Bertrand-like nature of
competition in the industry. 44
The patterns in the transition costs reflect the transition patterns for plant size found in
Table 2 on page 12. Entry costs are $1.0 million for small plants and $1.7 million for large
plants. This is in line with substantial differences in machinery and land for bigger plants.
There are also large costs of increasing the size of a plant. It takes about $0.3 million to
grow a plant from small to medium, $1.8 million to ramp it up from small to large, and $0.1
million to take a plant from medium to large. Thus it is cheaper to enter as a small plant and
grow to a large plant in the next period, and, indeed, 80% of plants enter as small plants.
Finally, the model also estimates substantial costs of ramping back down the size of a plant.
These large transition costs imply that plants have a weak response to demand shocks on
both the extensive (i.e., entry) and intensive (i.e., size) margin.
A bigger past size reduces the costs of growing a plant at some future point, and small
plants that were medium or large in the past find it easier to ramp back up. Likewise, a
medium-size plant that was large in the past has a lower cost of reverting back to being
a large plant. This dependence of transition costs on size in previous years lowers implicit
adjustment costs, since a plant can shrink today and retain the ability to cheaply increase
its size in the future.
6.1 Model Fit
To evaluate the fit of the model, I compare the evolution of the concrete industry to the
one predicted by the model. I obtain the model’s prediction by computing an equilibrium to
44If I remove market indicators from the covariates z in the auxiliary regression, I find substantially smallereffects of competition. Thus it is still important to target these market-level moments.
31
the dynamic game with a discrete action stochastic algorithm (DASA) presented in Online
Appendix A given the estimated parameters (henceforth θ) in Table 6. The DASA is an
adaptation of the Stochastic Algorithm of Pakes and McGuire (2001).45 This equilibrium
needs to be computed for all four market categories, since they have different demand Dµ
processes.
Using computed policies and the demand transition process, I simulate the model from
the observed states in 1976 until 1999.46 Table 7 shows moments for the data (Column I),
the simulation from the model’s prediction given estimates θ (Column II), and simulation
using the estimated CCP’s P µ (Column III).
The model does well at matching the distribution of plant size, as 47% of plants are small
in the data and versus 53% in the simulation, and 25% plants are large in the data, versus
24% in the simulation. However, the model under predicts the amount of turnover, as entry
and exit rates are 5.5% in the data, versus 2.9% in the simulation. Essentially, the model
under-predicts the amount of idiosyncratic shocks that yield entry and exit, rather than
demand driven turnover. Yet, the model does better at matching the frequency at which
firms change their sizes, predicting than 10% of plants increase their size and 9% reduce
their size each year, which replicates the rate at which plants grow and shrink in the data.
The model also predicts market structure quite accurately. There are, on average, two
plants per market in the data, and the model also forecasts 2.0 plants. Decomposing predicted
market structure, in the data one percent of markets have no plants, 45% are monopoly
markets, 27% are duopoly markets, and are 26% of markets have more than two plants. The
model predicts the same number of monopoly markets (43%) and slightly fewer markets with
45To compute counterfactual industry dynamics, I assume the existence of a single symmetric Markovperfect equilibria per market category µ. Besanko, Doraszelski, Kryukov, and Satterthwaite (2010) showthat this assumption may be problematic.
46Since the estimation of θ used year-to-year moments, rather than predictions on the entire time path ofthe industry, I am evaluating the model based on moments that were not used in estimation.
32
I II IIIMoments Real Data Simulated Data Simulated Data
(1976-1999) using Model θ∗ using CCP Pµ∗Plant Level MomentsShare of Small Plants 48% (1%) 53% (1%) 52% (1%)Share of Medium Plants 27% (0%) 23% (1%) 26% (1%)Share of Large Plants 25% (1%) 24% (1%) 22% (1%)Entry Rate 5.8% (0.0%) 2.9% (0.2%) 7.0% (0.2%)Exit Rate 5.4% (0.0%) 2.9% (0.2%) 7.3% (0.2%)Ramping Up Rate 10% (0.1%) 10% (0.3%) 11% (0.2%)Ramping Down Rate 9% (0.1%) 10% (0.5%) 9% (0.2%)
Market Level MomentsNumber of Plants Per Market 2.0 (0.2) 2.0 (0.4) 2.3 (0.1)No Plants in Market 2% (0%) 4% (1%) 18% (1%)Monopoly Market 46% (1%) 43% (1%) 32% (0%)Duopoly 26% (1%) 29% (1%) 18% (0%)More than 2 plants 26% (1%) 24% (1%) 32% (1%)
Number of Plants in Category 1 1.08 (0.00) 1.18 (0.02) 1.42 (0.06)Number of Plants in Category 2 1.76 (0.01) 1.80 (0.05) 2.20 (0.06)Number of Plants in Category 3 2.62 (0.01) 2.32 (0.08) 2.98 (0.08)Number of Plants in Category 4 4.15 (0.04) 4.25 (0.16) 4.44 (0.10)
Coefficient of VariationNumber of Plants within Market 0.7 (0.1) 0.6 (0.2) 1.7 (0.0)Correlation Demand and Plant Size 0.23 (0.01) 0.26 (0.01) 0.23 (0.01)Correlation Demandand Number of Plants 0.54 (0.01) 0.67 (0.01) 0.28 (0.01)
Note: ∗ Data is simulated using either II- computed policies given estimated parameters θ or III-
estimated conditional choice probabilities Pµ, taking as an initial condition the markets in 1976.
The market block bootstrapped standard errors shown in parenthesis use 50 bootstrap replications.
Table 7: Model Fit
33
more than two plants (24%). The model also does a good job at matching the number of
plants in each market category µ, even though the only way that market categories matter
is through differences in the estimated demand transition process Dµ.
To highlight the model’s ability to predict changes in the number of plants, I compute the
coefficient of variation (henceforth CV) of the number of plants within a market. The data
and the model predict a CV of 0.7 and 0.6 respectively. The correlation between market size
and the number of firms is 0.5 in the data, versus 0.7 in the model’s prediction. Likewise, the
correlation between market size and plant size (where plant size is just the integers 1, 2, and
3 corresponding to small, medium, and large) is 0.23, which is well matched by the model
(0.26). In sum, the model captures many features of the path of the ready-mix concrete
industry from the late 1970’s to 2000.
The 25-year forecasts using the CCP’s in Column III also match the path of the ready-
mix concrete industry. This precision is important, as the accuracy of the CCPII’s estimates
rely the ability of the estimated CCP’s to reproduce firms’s expectations about the evolution
of the industry. The distribution of plant size, as forecast by the CCP’s, much like the data,
with 52% small plants, 26% medium plants, and 22% large plants. Entry and exit rates are
somewhat higher - 7.1%, versus 5.6% in the data - while the rate at which plants either grow
or shrink is 9%, close to the data. Finally, the CCP’s forecast somewhat more plants in the
industry than what is indicated by the data - 2.3 plants per market, versus 2.0 in the data.
7 Counterfactual Industry Dynamics
There are substantial local fluctuations in construction activity. How do these demand shocks
affect the ready-mix concrete industry? The counterfactual that I consider would remove
much of the short-term fluctuation in construction activity at the county level.
Consider the policy where local governments allocate construction budgets to smooth
34
out changes in demand. Government commits to a five-year sequence of contracts so that
demand held remains constant over the next five years. Demand stays fixed at its five-year
expected level (given current demand), so on average, plants receive the same level of demand
over the next five years both with and without this policy. After five years are up, demand
reverts to the level it would have had absent demand smoothing, and the smoothing policy
is repeated. The long-run path of demand remains unchanged; this policy simply eliminates
short-run “wiggles” in demand.47,48
For illustrative purposes, I also show the effect of two other demand-smoothing policies.
The first is constant demand. Second, to directly investigate the effect of demand smoothing
on the equilibrium strategies used by firms, such as how responsive they are to demand
shocks, I consider “myopic” firms. These firms believe that demand is constant over time,
but in fact, demand evolves following the process estimated in the data (Dµ).
Demand smoothing may alter the rate at which firms both enter and exit, as well as
how often firms ramp up and shrink down their size. Moreover, removing fluctuations may
change the stationary distribution of the industry, so I also look at the effect of the demand-
smoothing policy on the size distribution and market structure of the ready-mix concrete
industry.
47This policy would be fairly easy to implement, since it simply relies on local governments being able toborrow and save over relatively short periods of time and assumes that construction projects, such as roads,are efficiently broken up across years.
48More formally, I compute the smoothed demand level S as a function of initial demand M0 = M as
S(M) =1
5
4∑t=0
ED[M t|M0 = M
]And the smoothed process is fixed in periods t in which t mod 5 6= 0 and evolved in periods where tmod 5 = 0 via:
M ∼ D5(M t|M0)
where D5 is the five year transition process generated by repeating the one year transition process D for fiveyears.
35
7.1 Demand Smoothing, Turnover, and Size Changes
Table 8 shows statistics on entry, exit, and size changing in the ready-mix concrete industry
for the four different demand processes I consider. I present annual statistics 25 years after the
policy has been put into place to allow the industry to adjust to the new demand process.49
Unsmoothed 5 Years of Constant Firms believe
Demand (Dµ) Smoothing Demand demand is constant
TurnoverEntry Rate 2.7% 2.2% 2.2% 4.1%Exit Rate 2.9% 2.0% 2.1% 4.5%Change in Size Rate 20% 18% 17% 18%
InvestmentSunk Entry Costs 132 137 107 155per year(in million $)Size Changing Costs 307 496 407 337per year (in million $)Total Plants 3,643 5,433 4,264 3,879
Table 8: Demand Smoothing, Turnover and Size Changing
The five-year demand-smoothing policy reduces turnover on the entry/exit margin, and
yet has little effect on the frequency at which firms change their size. The turnover rate falls
by 25% – from 3.0% in the unsmoothed case to 2.2% with five-year smoothing. The rate at
which firms change their size is approximately the same: 17% versus 20% in the unsmoothed
case. Moreover, even when all demand changes are eliminated, the turnover and the size
changing rates are similar to those where demand is smoothed over only five-year periods.
Meddling with the demand process would reduce – but not eliminate – turnover. This
is consistent with both the descriptive work of Dunne, Roberts, and Samuelson (1988) and
the fact that the entry and exit rates (per incumbent plant) are virtually uncorrelated at
49These annual averages are computed for years 25 to 50 after the policy has been put into place. Thereis additional exit and entry in the first 25 years as the market adjusts to its new stationary distribution, butthese are relatively few in number, compared with overall turnover. I use the effect 25 years after the policyhas been put into place to separate the dynamics of transitioning to a new stationary distribution from theentry and exit patterns in the stationary distribution itself.
36
the county-year level.50 These indicate that turnover is not generated solely by market-level
shocks, which would lead to either entry or exit - not both, but rather, by idiosyncratic
shocks εti, such as productivity shocks.51 Remembering that the model under-predicts the
amount of turnover due to idiosyncratic shocks, this 25% decrease in turnover most likely
overstates the effect of reducing demand volatility on turnover.
Yet, this explanation is incomplete, since it ignores firms anticipation of changes in de-
mand. Indeed, increased demand volatility makes firms less sensitive to demand shocks. As
an example, industry dynamics with either constant or i.i.d. demand are identical, as both
demand processes imply that a firm’s expected level of demand never changes, even though
one demand process has no volatility, and the other one is highly volatile. If I take firms
that use the policies corresponding to a constant level of demand but subject them to the
demand process estimated in the data (the myopic counterfactual), I find that the turnover
rate would increase by 50% to 4.3% per year, and the rate at which firms change their size
would remain at 16% per year. Moreover, I find that firms become far more sensitive to
variation in demand. A market-fixed effect regression of the number of plants on log con-
struction employment yields a coefficient of 0.16 when firms use policies corresponding to
variable demand, versus 0.30 when their policies correspond to constant demand. Thus the
expectations of future changes in demand blunt the response to current demand shocks, and
this is why we see such a small reaction, in terms of turnover and investment, to demand
changes.
Using the estimates of the model in Table 6, I find that in the base case, sunk entry costs
are $132 million dollars per year, and transition costs are $307 million per year. With the
demand-smoothing policy, these costs rise to $137 million per year of sunk entry costs, and
50Dunne, Roberts, and Samuelson (1988) show that entry and exit rates are highly correlated at theindustry level, while I show that in the ready-mix concrete industry these are uncorrelated at the county-year level.
51See Collard-Wexler (2009) for the importance of productivity volatility in this industry.
37
$496 million per year of size changing costs.
This 44% increase in investment is almost entirely due to the 39% increase in the number
of plants in the industry under the demand-smoothing policy. In short, plants invest at the
same rate, but there are more of them.
7.2 Demand Smoothing and Industry Composition
Table 9 shows the effects of the demand-smoothing policy on the number of plants in the
industry, fixed costs, plant size, and the industry’s market structure. As in Table 8, I show
annual averages for the industry between 25 and 50 years after the policy is put into place.
I will compare constant demand to the unsmoothed case and then look at the effects of the
five-year smoothing policy.
Unsmoothed Constant 5 YearsDemand Demand of Smoothing
Total Plants 3,645 4,264 5,433Fixed Costs 717 878 1,109(per period in millions of $)
Industry CompositionSmall Plants 54% 48% 49%Medium Plants 23% 23% 24%Big Plants 23% 29% 28%
Market StructureMarkets with no plants 5% 8% 1%Markets with 1 plant 43% 36% 25%Market with 2 plants 28% 24% 29%Markets with more than 2 plants 25% 32% 46%
Table 9: Demand Smoothing and Industry Composition
38
7.2.1 Constant Demand
The constant demand process predicts sixteen percent more plants than the unsmoothed
demand process (4,264 versus 3,645). Market structure also differs, as the number of plants
per market is more dispersed under constant demand, with 8% of markets having no plants,
versus 5% in the unsmoothed case, and 32% of markets having more than 2 plants, versus 25%
in the unsmoothed case. Constant demand spreads out the distribution of the net present
value of demand, as a market with high demand will have high demand forever, and likewise,
markets that have low demand retain it in perpetuity.
This difference in the cross-sectional distribution of demand also changes the industry’s
plant size distribution, as 29% of plants are large under constant demand, versus 23% under
unsmoothed demand. Figure 3 showed that bigger markets have larger plants. Thus a change
in the distribution of market size alters the industry’s plant size distribution.52
7.2.2 Five Years of Smoothing
Since demand is very volatile and shocks are short-lived, removing five-year changes in de-
mand has a large effect on the inter-temporal variance of demand. This policy increases
the number of plants from 3,645 to 5,433 and raises fixed costs from $717 million to $1,109
million per year.
If profits per consumer are either increasing or decreasing with demand (holding market
structure fixed), then period profits are either a convex or concave function of demand. This
has important implications on the effect of smoothing demand volatility. With a concave
profit function (with respect to demand), by Jensen’s Inequality, less inter-temporal volatility
of demand raises the expected profitability of a market. I call this the “market expansion
effect” of demand smoothing.
52 I find more large plants, even though in principle I could find either more small plants or more bigplants given the distribution of demand in different markets.
39
Figure 3 showed that the relationship between the number of plants in a market and
construction demand is concave. This implies that profits per consumer are decreasing with
demand.53 I speculate that this effect is due to congestion costs for concrete deliveries when
demand is particularly high. There are greater costs involved in making multiple deliveries
because of labor and machinery shortages or because some deliveries cannot be made during
the weeks of the year when demand peaks. Congestion is more likely when yearly demand is
higher.
In contrast, in other industries reducing demand volatility might lower the number of
firms in the market. For instance, in the market for electric power, short-run profits are a
convex function of short-run demand because aggregate bid curves typically take a highly
convex “hockey stick” shape.54 Thus with lower demand volatility one might expect fewer
power plants to be built.
The “market expansion effect” for the ready-mix concrete industry increases a market’s
profitability, and this raises the number of firms at which the free-entry condition binds.
The number of markets served by more than one plant increases from 52% to 74%. As well,
the increase in market size generated by the demand-smoothing policy’s “market expansion
effect” raises the share of large plants from 23% to 28%, while the share of small plants goes
down from 54% to 49%, and the share of medium plants stays about the same.
Note that the five year demand-smoothing policy has very different effects than the
constant demand policy: It only reduces the inter-temporal variance of demand, while the
constant demand policy also increases the cross-sectional dispersion of demand. These two
effects make it difficult to untangle the effects of a constant demand policy.
53This implication can be tested by estimating the elasticity of profits (measured by sales minus all inputcosts) with respect to construction employment. I find a very inelastic response to construction employmentwhen running this regression. However, attenuation bias is also a plausible explanation for this estimate.
54See, for instance, the aggregate bid curves in Figure 2-1 on page 34 of http://www.
monitoringanalytics.com/reports/PJM_State_of_the_Market/2010/2010-som-pjm-volume2-sec2.
pdf.
40
7.3 Consumer and Producer Surplus
Table 10 summarizes the differences in welfare between the five-year demand-smoothing
policy and unsmoothed demand. I show these effects on the net present value of surplus
(henceforth, NPV) for consumers, incumbents, and potential entrants.
Change in Net Present Value of
Consumer Surplus $860 MillionProducer Surplus for Incumbents -$609 MillionProducer Surplus for Potential Entrants -$36 Billion
Note: Numbers in this table refer to the difference in the net present value of surplus (using a 5%
discount rate) between five years of smoothing and unsmoothed demand, averaged between 25 and
50 years after the policies are put into place, using 1976 as an initial states.
Table 10: Welfare Effects of Demand-Smoothing Policies
For the 19% of markets that were formerly monopoly markets, but became competitive,
prices would fall by 3%, based on the estimates in Figure 2. Taking into account all changes
in market structure, decreases in price due to additional competition (holding purchases
of concrete fixed) transfer 43 million dollars per year, or 860 million dollars in NPV, from
producers to consumers. This number is a lower bound on the increase in consumer surplus,
as any elasticity in the demand for concrete would add to it.
In an oligopoly model with free entry, it is not clear how an increase in the number of
firms in a market will affect producer welfare. I find that producer surplus for incumbents
would decrease from 3.3 to 2.7 billion dollars in NPV under the demand-smoothing policy,
representing a 20% fall.55 I also compute producer surplus for potential entrants, who repre-
55 To compute producer surplus, I reformulate the problem in terms of choice-specific value functions.Thus producer surplus is just:
PS =∑
i is incumbent
V i(xi0) +
∞∑t=0
βt∑
i is entrant
V i(xit) (14)
which is just the ex-ante value function for incumbents, plus the discounted value of entrants in the future,which needs to be monitored since I assume that if an entrant does not enter they get a continuation value
41
sent 80% of the “firms” in the data. Their surplus falls from 134 to 98 billion dollars in NPV
when the demand-smoothing policy is implemented, a 31% decrease. However, the numbers
for the surplus of potential entrants are suspect, since the vast majority of this surplus is
derived from 98.7% of potential entrants who choose never to enter, yet receive a payoff from
their private information shock εa0 . Surplus from firms that do not enter is truly an artifact
of the model, since how do we interpret the profits of firms that choose not to enter?56
8 Conclusion
Due to the turbulence of local construction markets, the ready-mix concrete industry is
subject to large fluctuations in demand. These fluctuations have substantial effects on the
composition, size, and investment level in the industry. Specifically, I considered a policy in
which the government would sequence its construction budgets in such a way as to eliminate
five-year changes in demand, while retaining longer-run movements.
I estimated an oligopoly model of entry/exit and discrete investment and used it to eval-
uate the industry’s response to this policy. This model allowed for considerable heterogeneity
between plants, in terms of current and past size, as well as persistent differences between
local markets.
Demand need not have a linear effect on plant profits, especially if marginal costs increase
with the number of concrete deliveries. Absent linearity, any changes in the volatility of de-
mand affect the profitability of a market, and hence, the number of plants it can support. For
this industry, a reduction in inter-temporal volatility of demand has a “market expansion”
effect. This effect is similar to an increase in market size and raises both the number and
of 0. The ex-ante value function is V (x) =∑j∈Ai
W (j|x)Ψ(j|x) + γ −∑j∈Ai
ln (Ψ(j|s)) Ψ(j|s).56Potential entrants represent more than 80% of the players in the game, and potential entrants who choose
to enter are 1.3% of all potential entrants, as is illustrated in the first row of Table 2 on page 12. Thesefigures rely crucially on the assumption that there are 10 firms in each market, so the number of potentialentrants are given mechanically as 10 minus the number of incumbents.
42
average size of plants in the industry.
This demand-smoothing policy lowers the amount of turnover in the industry by 25%,
but leaves the rate at which firms change their size unaffected. In this industry, large sunk-
and size-changing costs make it expensive to respond to demand shocks. Furthermore, firms
are unlikely to react to demand shocks when demand is very volatile, since these demand
shocks convey little information on future profitability.
High volatility of plant-level demand, and associated plant-level profitability, is a feature
of many industries. As Collard-Wexler, Asker, and De Loecker (2011) show, the volatility of
plant profitability, driven in part by demand, differs substantially across similar industries
in various countries. This paper indicates that the consequences of volatility may not be ex-
pressed by higher turnover or more volatile investment. Instead, demand volatility transforms
the size and structure of the industry.
ReferencesAguiregabiria, V., and P. Mira (2007): “Sequential Estimation of Dynamic Games,” Econo-
metrica, 75(1), 1–54.
Auerbach, A., W. Gale, and B. Harris (2010): “Activist fiscal policy,” The Journal of Eco-nomic Perspectives, 24(4), 141–163.
Bajari, P., L. Benkard, and J. Levin (2007): “Estimating Dynamic Models of ImperfectCompetition,” Econometrica, 75(5), 1331–1370.
Besanko, D., U. Doraszelski, Y. Kryukov, and M. Satterthwaite (2010): “Learning-By-Doing, Organizational Forgetting, and Industry Dynamics,” Econometrica, 78(2), 453–508.
Bresnahan, T. F., and P. C. Reiss (1991): “Entry and Competition in Concentrated Markets,”Journal of Political Economy, 99(5), 33.
Campbell, J., and H. Hopenhayn (2005): “Market Size Matters,” The Journal of IndustrialEconomics, 53(1), 1–25.
Collard-Wexler, A. (2009): “Productivity Dispersion and Plant Selection in the Ready-MixConcrete Industry,” Working Paper, New York University.
Collard-Wexler, A., J. Asker, and J. De Loecker (2011): “Productivity Volatility and theMisallocation of Ressources in Develloping Economies,” Working Paper, New York University.
Doraszelski, U., and M. Satterthwaite (2010): “Computable Markov-Perfect Industry Dy-namics,” RAND Journal of Economics, 41(2), 215–243.
43
Dunne, T., S. D. Klimek, M. J. Roberts, and D. Xu (2006): “Entry and Exit in GeographicMarkets,” Working Paper, Penn State University.
Dunne, T., M. J. Roberts, and L. Samuelson (1988): “Patterns of Firm Entry and Exit inU.S. Manufacturing Industries,” RAND Journal of Economics, 19(4), 495–515.
Ericson, R., and A. Pakes (1995): “Markov-Perfect Industry Dynamics: A Framework for Em-pirical Work,” The Review of Economic Studies, 62(1), 53–82.
Fershtman, C., and A. Pakes (2012): “Dynamics Games with Asymmetric Information: AFramework for Empirical Work,” Quarterly Journal of Economics.
Gourieroux, C., and A. Monfort (1996): Simulation-based econometric methods. Oxford Uni-versity Press, USA.
Gourieroux, C., and A. Montfort (1993): “Indirect inference,” Journal of Applied Economet-rics, 8, S85–S118.
Gowrisankaran, G. (1999): “Efficient Representation of State Spaces for Some Dynamic Mod-els,” Journal of Economic Dynamics and Control, 23(8), 1077–1098.
Griliches, Z., and J. Hausmann (1986): “Errors in Variables in Panel Data,” Journal of Econo-metrics, 31, 93–118.
Hotz, V. J., and R. A. Miller (1993): “Conditional Choice Probabilities and the Estimationof Dynamic Models,” The Review of Economic Studies, 60(3), 497–529.
Hotz, V. J., R. A. Miller, S. Sanders, and J. Smith (1994): “A Simulation Estimator forDynamic Models of Discrete Choice,” The Review of Economic Studies, 61(2), 265–289.
Jarmin, R. S., and J. Miranda (2002): “The Longitudinal Business Database,” Working Paper,Center for Economic Studies, US Census Bureau.
Keane, M., and A. Smith Jr (2003): “Generalized Indirect Inference for Discrete Choice Models,”Working Paper, Yale University.
Kosmatka, S. H., B. Kerkhoff, and W. C. Panarese (2002): Design and control of concretemixtures, Engineering bulletin ; 001. Portland Cement Association, Skokie, Ill., 14th edn.
Manuszak, M., and C. Moul (2008): “Prices and Endogenous Market Structure in Office SupplySuperstores,” Journal of Industrial Economics, LVI(1), 94–112.
Pakes, A., S. Berry, and M. Ostrovsky (2007): “Simple estimators for the parameters ofdiscrete dynamic games, with entry/exit examples,,” RAND Journal of Economics, 38(2), 373–399.
Pakes, A., and P. McGuire (2001): “Stochastic Algorithms, Symmetric Markov Perfect Equi-librium, and the ’Curse’ of Dimensionality,” Econometrica, 69(5), 1261–1281.
Pesendorfer, M., and P. Schmidt-Dengler (2008): “Asymptotic Least Squares Estimatorsfor Dynamic Games,” The Review of Economic Studies, 75, 901–928.
Powell, W. B. (2007): Approximate Dynamic Programming: Solving the Curse of Dimensionality,Wiley Series in Probability and Statistics. John Wiley and Sons.
44
Ryan, S. P. (2012): “The Costs of Environmental Regulation in a Concentrated Industry,” Econo-metrica.
Syverson, C. (2004): “Market Structure and Productivity: A Concrete Example,” Journal ofPolitical Economy, 112(6), 1181–1222.
45